Description
Political discourse distributes responsibility through both actor mention and the event roles and predicate environments assigned to actors. This study develops an auditable actor-role-event framework and applies it to 156 Donald Trump speeches (June 2015–July 2026; 843,837 lexical tokens). The frozen analysis contained 48,375 primary actor mentions and 1,858 actor-event frames. Because the original row-level export was unavailable, the frames were regenerated from frozen corpus text using transparent v2.5 rules and constrained by 39 archived aggregate invariants, all reproduced exactly. A two-coder human audit of 400 candidates retained 382 frames, yielding 95.5% PPV (95% CI [93.0%, 97.1%]); actor, role, and event labels were 100% accurate conditional on human-gold inclusion. In the regenerated ledger, 88.0% of I/me, 87.4% of we/us, and 94.1% of administration frames were agentive; we/us did not differ from I/me (OR = 1.02). Foreign-state actors and political opponents showed elevated odds of DESTRUCTIVE_HARM versus CONSTRUCTIVE/PROTECTIVE predicates (ORs = 17.55 and 6.19), robust across 21 leave-one-predicate-out refits. Human-gold role accuracy was 89.5% for a subject proxy, 99.5% for a syntactic baseline, and 100% for the normalized system. Extraction recall remains unidentified. Speech-equal weighting preserved role ordering, although the political-opponent contrast attenuated; non-extracted mentions were not sampled for recall estimation.
Thông tin các tác giả
Nguyen Ngoc Vu
Ho Chi Minh City University of Foreign Languages – Information Technology
vunn@huflit.edu.vn
Từ khóa
political discourse; social actors; agency; semantic roles; responsibility attribution; corpus-assisted discourse studies; computational linguistics; Donald Trump
